Evolutionary Computation for Econometric Modeling

This paper shows that, in case of high sensitivity to data econometric modeling, using evolutionary algorithms (Genetic Algorithms GA and Simulated Annealing SA) is better than using classical gradient techniques. The evaluation of the algorithms involved was performed on a short form of an economic macromodel. The optimization task is the model’s solution, as function of the initial values (in the first stage) and of the objective functions (in the second stage). We proved that a priori information help “elitist” algorithms (like SA) to obtain best results; on the other hand, when one has equal believe concerning the choice among different objective functions, GA gives a straight answer. Analyzing the average related bias of the model’s solution proved the efficiency of the stochastic optimization methods presented. Introduction This paper is a study on the initial values’ influence on finding the optimal solution for some economic macromodel. Empirical trials showed the high sensitivity of the model’s output to small variations of the input values. A second goal of our approach consists in rising some questions on the adequacy of different objective functions for the same econometric model. The analysis of the suitable optimization techniques is depicted in Section 1. The numerical results of applying these techniques for finding the extreme global solution of the macromodel, including a comparative discussion on the several objective functions that one can associate to the macromodel, are presented in Section 2. We omit the concrete form of the econometric model, as irrelevant. 1 Evolutionary Computation for Function Optimization When initiating a numerical analysis of some econometric model, one has to face a difficult problem: What optimization algorithm should he/she use? Usually, this problem is solved by a reduction mechanism: classical algorithms from Operation Research (SIMPLEX, e.g.) are excluded, as they are limited to convex regular functions. Gradient techniques (Gauss-Seidel, e.g.) are the first candidates, as their speed and precision on problems with a single local optima is highly appreciated. Probabilistic search algorithms (like GA – Michalewicz (1994) or SA – Laarhoven et al. (1987)) come next: they are easy to